FinSeer

Retrieval-augmented Large Language Models for Financial Time Series Forecasting

Paper · Models

Overview

This repository contains the code for FinSeer, a domain-specific retriever for financial time-series forecasting, used in a retrieval-augmented generation (RAG) framework for stock movement prediction. FinSeer is trained with candidate selection refined by LLM feedback and a similarity-driven objective, and the retrieved sequences are fed into StockLLM, an LLM fine-tuned for stock movement prediction. The code covers indicator calculation, retriever training, and stock movement prediction with and without retrieval.

Resources on Hugging Face

Resource	Type	Description
FinSeer collection	Collection	All FinSeer resources
TheFinAI/FinSeer	Model	Our retriever
TheFinAI/StockLLM	Model	Our fine-tuned stock LLM

Repository Structure

src/
├── 1_calulate_indicators/          # financial indicator calculation
├── 2_train_retriever/              # LLM feedback scoring and candidate selection
└── 3_stock_movement_prediction/    # embeddings, similarity, and RAG prediction

Installation

# for baseline RAG models and retriever training
pip install InstructorEmbedding
pip install -U FlagEmbedding
pip install sentence-transformers==2.2.2
pip install protobuf==3.20.0
pip install yahoo-finance
python -m pip install -U angle-emb
pip install transformers==4.33.2  # UAE

Train the Retriever

Step 1. Get LLM feedback scores (src/2_train_retriever/get_llm_feedback_scores.py)

• dataset: acl18, bigdata22, stock23
• target: the file to save, a JSON with LLM probability scores

parser = argparse.ArgumentParser(description='test')
parser.add_argument('--dataset', default='acl18', type=str)
parser.add_argument('--target', default='acl18.scored.json', type=str)
args = parser.parse_args()
get_all_scores(llm='StockLLM')  # llama family are all supported for this code

Step 2. Select positive and negative candidates (src/2_train_retriever/select_positive_and_combine.py)

Before this step, you should have generated acl18.scored.json, bigdata22.scored.json and stock23.scored.json.

This step selects candidates for all three datasets and generates a combined train.scored.json file.

Then, follow the steps in this link to fine-tune your own FinSeer using the train.scored.json data.

Predict Stock Movement with the RAG Model

Step 1. Get embeddings of queries and candidates (src/3_stock_movement_prediction/get_embeddings.py)

• q_or_c: query or candidate; we generate the embeddings of query sequences and candidate sequences separately

parser = argparse.ArgumentParser(description='test')
parser.add_argument('--test_dataset', default='bigdata22', type=str)
parser.add_argument('--embedding_model', default='e5',
                    choices=['instructor', 'uae', 'bge', 'llm_embedder', 'e5', 'FinSeer'])
parser.add_argument('--q_or_c', default='candidate')
args = parser.parse_args()

Step 2. Calculate the similarity of queries and qualified candidates (and get the top-5 related candidates).

Step 3. Predict stock movement with or without retrieval. This is what our three files do:

• 1_no_retrieval.py
• 2_random_retrieval.py
• 3_similarity_retrieval.py

Citation

If you find FinSeer useful, please cite:

@misc{xiao2025retrievalaugmented,
  title={Retrieval-augmented Large Language Models for Financial Time Series Forecasting},
  author={Mengxi Xiao and Zihao Jiang and Lingfei Qian and Zhengyu Chen and Yueru He and Yijing Xu and Yuecheng Jiang and Dong Li and Ruey-Ling Weng and Min Peng and Jimin Huang and Sophia Ananiadou and Qianqian Xie},
  year={2025},
  eprint={2502.05878},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}

License

The code in this repository is released under the MIT License. Datasets and models on Hugging Face keep their own licenses, stated on each card.

---

Built by The Fin AI · Hugging Face · GitHub